AI Diffusion and the Possibility of Full Unemployment: An Urgent but Neglected Issue

Abstract

Artificial Intelligence (AI) is often viewed as a potential catalyst for mass unemployment. However, there has been a notable lack of contributions published in economics journals on this topic, with virtually none exploring the possibility of full unemployment. This article has three main goals. First, it defines full unemployment and outlines its key theoretical features. Second, it seeks to explain economists’ reluctance to acknowledge AI’s potential to induce mass unemployment and, a fortiori, full unemployment. Finally, the article aims to clarify why, despite its low probability, full unemployment deserves serious theoretical consideration and detailed analysis by economists. It suggests that full unemployment might be seen as the long-term asymptotic endpoint of short-term dynamics marked by rising unemployment, even if the actual magnitudes never reach it.

JEL Classification: E24, J23, J64, O33

Keywords

technological unemployment full unemployment mass unemployment artificial intelligence AI fourth industrial revolution long-period positions

1. Introduction

The concept of “full unemployment” is entirely new to theoretical economic studies. The term was popularized by the futurist scholar James Dator (see, e.g., Dator 2016, 2020), re-proposed in D’Orlando and Ferrante (2022: 46–47), and further developed in D’Orlando (2024). Korinek and Juelfs (2022) refer to a similar situation as “strong economic redundancy of labor.” It can be defined in various ways, but, in essence, it describes either a situation in which the impact of artificial intelligence (AI) on the labor market is such that no human worker is willing to work at the equilibrium wage rate, since the latter is below subsistence, or one in which no firm is willing to hire workers at a wage rate above subsistence. Although full unemployment, defined as zero employment across all industries, can be considered only a long-period theoretical position toward which the economic system tends asymptotically to converge over time, and mass unemployment might be a more realistic outcome, economists have not only ignored full unemployment but have also given mass unemployment little space in their studies: Indeed, most mainstream models still rely on the very different concept of full-employment equilibria. In this article, it is argued that, on the contrary, after the Fourth Industrial Revolution and the diffusion of AI, economists should consider the possibility that technological progress might push the economy in the long run toward full unemployment rather than toward full employment, regardless of whether this outcome will actually be reached.

Until a couple of years ago, the consensus among scholars, including both economists and scholars from other disciplines, was that contemporary technological progress could negatively affect employment opportunities and wages for low-skilled workers or those performing routine tasks, resulting in greater income inequality, but could not significantly affect long-term aggregate unemployment. Moreover, nobody could even imagine that technological progress might ever make “machines” capable of substituting for humans in intellectual, artistic, and, in general, non-routine tasks.

Nowadays, everything has suddenly changed. ChatGPT, Claude, Gemini, DALL-E, and other AI tools are just a few examples of how AI and machines endowed with AI are already impacting most tasks and jobs, both manual and intellectual, both routine and non-routine, and even artistic.

Unsurprisingly, traditional and less traditional media, scholars across various disciplines, and even politicians have immediately recognized that the diffusion of AI and machines endowed with it may have a massive impact on the labor market. Nonetheless, there are very few contributions published by economists in economics journals suggesting that AI can generate technological mass and, a fortiori, full unemployment. The few studies explicitly addressing mass and full unemployment often appeared as unpublished papers, working papers, books for the general audience, and book chapters: Only in a couple of cases have they been published in scientific journals, indicating a certain reluctance among journal editors and reviewers to admit even the possibility that mass or full unemployment might be the object of serious theoretical studies.

So, the question arises: Why did economists, with few exceptions, underestimate the impact of AI on employment? Is this a consequence of biases, or is their attitude well-grounded?

The first possible explanation for economists’ attitude may be an overestimation of the role of empirical data: because the bulk of empirical research supports the idea that AI has had, until today, a mild impact on unemployment, mass and full unemployment are considered not worth theoretical interest.

The second possible explanation may be the a priori acceptance of the continued validity of an old theory, the so-called “compensation theory,” according to which exogenous shocks, such as technological progress, may well divert the system from full employment, generating waves of short-term and/or sectoral unemployment, but many compensatory forces (from wage flexibility to increased demand) would automatically bring the system back to full employment.

We argue that both arguments are flawed. Today’s empirical data are uncorrelated with the future impact of AI on unemployment because AI’s diffusion into productive processes is in its earliest stages, so today’s empirical evidence can neither prove nor disprove anything. The compensation theory is founded on the implicit assumption, which until today had been quite reasonable, that substituting human workers entirely or almost entirely with machines was impossible (that the substitution rate between humans and machines would never equal one), a circumstance that is now at least disputable.

However, a third possible explanation for economists’ attitude exists. This explanation is rooted in the dominance of the mainstream (neoclassical) view, which holds that the economic system tends spontaneously to reach full-employment equilibria, characterized by the absence of involuntary unemployment. If mainstream economists cannot conceive the existence of involuntary unemployment equilibria, they a fortiori cannot even imagine that the system might converge to full unemployment, a situation that, moreover, involves phenomena that are difficult for traditional theory to handle, such as an income distribution explained by class struggle rather than by factor endowments and marginal productivity.

In this article, it is argued that contrary to the prevailing view among economists, the concept of full unemployment warrants theoretical interest. I argue that, paradoxically, it is worthy of attracting theoretical interest even if the aim is to demonstrate that it cannot occur: One cannot simply assume this conclusion, one has to prove it. And to furnish such a proof, theoretical studies on the topic are unavoidable. In doing so, we should consider full unemployment as a tendential long-term outcome, similar to the long-period positions of classical-type, toward which actual, more realistic outcomes, such as mass unemployment, might asymptotically converge.

The article is organized as follows: Section 2 defines the concept of full unemployment and explains how the diffusion of AI during the Fourth Industrial Revolution might, for the first time, make this concept theoretically worthy of attention, regardless of its short-term actual likelihood. Section 3 examines possible explanations for economists’ belief that mass, and a fortiori full, technological unemployment are not concepts worthy of theoretical interest, focusing in particular on the still-relevant theory of compensation, the role of empirical evidence, and the dominance of the mainstream full-employment equilibrium approach. Section 4 questions the validity of the foundations of economists’ attitude toward full and mass unemployment described in section 2. Section 5 concludes by discussing whether the concept of full unemployment warrants theoretical interest, especially by examining whether it is legitimate to study phenomena that may or may not occur in the future. It also examines the possibility of viewing full unemployment as an attractor—a long-period position toward which reality could asymptotically tend in the absence of technological, political, or social obstacles — regardless of its likelihood in economic reality.

2. What Is Full Unemployment, and Why Should It Attract Theoretical Attention in the AI Era?

In the introduction, we have seen that full unemployment can be defined as a situation in which no human worker accepts a job offer, since the equilibrium wage is below subsistence (and therefore below all human workers’ reservation wages), or one in which no firm is willing to hire workers paying them a wage above subsistence. This “pure” theoretical definition of full unemployment implies that no human worker is employed in the entire economy. However, a more practical definition might treat full unemployment as a long-period position toward which the economic system tends to converge asymptotically, even if the actual reality remains characterized by mass, and possibly rising, unemployment.

Full unemployment, in any case, becomes relevant only in the context of the Fourth (or Fifth) Industrial Revolution, particularly after the diffusion of AI.

The term “Fourth Industrial Revolution” was introduced into theoretical debate by Schwab (2016), who refers to it as a phenomenon that “began at the turn of this century and builds on the digital revolution. It is characterized by a much more ubiquitous and mobile internet, by smaller and more powerful sensors that have become cheaper, and by AI and machine learning” (Schwab 2016: 11). The Fourth Industrial Revolution presents many distinctive features compared with past industrial revolutions. At the outset, these features were not fully recognized, not even by Schwab himself: Many scholars thought they were simply witnessing an evolution of the Third Industrial Revolution, since most of the development of AI was still to come, and the primary attention was on robots and further mechanization of the productive process, with AI mentioned only as one element of a longer list. In particular, in those days, robotization was considered a significant technological advance, with enormous implications for economic outcomes and income distribution (primarily through wage polarization and rising income inequality). As a result, many scholars regarded this process and/or these outcomes as key elements of the Fourth Industrial Revolution (see, e.g., Schwab 2016; Morrar et al. 2017; Xu et al. 2018). Other scholars began to recognize AI’s relevance, and some considered it the defining feature of a Fifth Industrial Revolution, distinct from mere robotization (see, e.g., Muir 2018; Mourtzis 2021; Ali et al. 2022). However, only in the last couple of years, with a greater understanding of AI’s potential, have some economists begun recognizing the enormous differences this revolution presents compared to previous ones (see, e.g., Acemoglu and Johnson 2023; Luo 2023; Korinek and Jeulfs 2022).

In particular, the diffusion of AI introduced a significant paradigm shift in the debate on technological unemployment: Introducing the steam engine, computers, and even robots into the productive process is utterly different from introducing AI and machines endowed with AI. During past industrial revolutions, unemployment could result from the fact that machines cooperated with human workers, raising labor productivity and thereby reducing the labor requirement for a given amount of production. In the Fourth Industrial Revolution, machines and AI might no longer cooperate with human workers to boost productivity but instead could replace them in most jobs and tasks, directly displacing human employment¹—a circumstance that was not even conceivable in the past.

Despite its early stage of diffusion, we can already today observe the first, not yet fully developed, peculiar consequences of the Fourth Industrial Revolution for the labor market. If the Third Industrial Revolution, albeit based on similar elements (i.e., Information and Communication Technologies and robotization), mainly affected manual workers and/or those performing routine tasks, AI appears capable of affecting a greater number of jobs and tasks, including creative and intellectual ones. Particularly relevant for this outcome has been the impact of Generative AI, a subset of AI capable of autonomously creating original content, including text, images, code, and audio. The most famous contemporary examples of Generative AI are Large Language Models like ChatGPT, Claude, and Gemini. Human beings must indeed drive these tools if they aim for high-quality results, but AI learns very quickly, so in the coming few years, human assistance might become superfluous. The same considerations can be extended to the figurative arts, since tools like Midjourney and Nano Banana are already capable of producing higher-quality artistic images than those produced by the average painter or photographer. High-skilled and non-routine jobs are at the same risk in almost all sectors of the economy: AI tools like Aidoc or Qure.ai are already capable of producing diagnoses of illness and therapies better than skilled physicians; other AI tools, like Synthesia or HeyGen, can generate avatars of virtual professors, transforming a written text into a multilingual video lesson, and tools like Coursera Coach can answer the questions of students; and so on. Therefore, even if mass substitution has not yet fully manifested in all these areas, it appears to be only a matter of time. Research indicates that tasks exposed to automation through the introduction of AI in the production process pose a risk to a significant portion of the workforce. These studies do not predict a complete replacement, a substitution rate equal to one; nonetheless, they suggest that up to half of the workforce—especially, but not only, those engaged in cognitive-task-oriented jobs—are already at risk of having their tasks performed by AI (see, e.g., Cazzaniga et al. 2024; Eloundou et al. 2024; Webb 2020). And since these studies focus on the impact of today’s AI on today’s jobs rather than tomorrow’s AI on tomorrow’s jobs, this concern will only increase as AI continues to evolve and spread.

In such a context, we inevitably have to ask ourselves what would happen if a mass replacement of human workers began and accelerated, generating waves of high unemployment. We can easily argue that, in this situation, it should be considered the possibility that the system, in the absence of automatic compensatory forces or adequate policies, might converge toward full unemployment, just as mainstream economists claim that, under the assumption of free competitive markets, the system tends toward full employment. Therefore, the asymptotically stable fixed point of this dynamic deserves theoretical attention.

To discuss the characteristics of this attractor, we can build a thought experiment based on the hypothesis that firms can choose whether to hire androids or human workers to produce goods and services. We call “androids” humanoid machines endowed with AI that can perform the same tasks as a human worker, but with higher productivity. The choice of referring to androids rather than simply to machines endowed with AI greatly simplifies the analysis without affecting its main results. We can imagine that androids are rented by their owners (i.e., capitalists) to entrepreneurs (i.e., firms). In the spirit of our thought experiment, assume also that (a) androids can substitute for workers in all jobs and tasks, and (s) androids’ productivity-to-remuneration ratio is higher than humans’ productivity-to-wage ratio in all sectors of the economy (for a wage rate above subsistence).

In formal mainstream economic terms, given the produced quantities, firms will hire androids (A) instead of human workers (L) if the value of androids’ marginal product $\bar{p .} \frac{\partial Q}{\partial A}$ divided by androids’ rent price (or remuneration: the price for the service of the factor, i.e., capitalists’ profits) $w_{A}$ is higher than the value of labor marginal product $\bar{p .} \frac{\partial Q}{\partial L}$ divided by workers’ wage rate $w_{L}$ , that is, if:

\frac{\bar{p .} \frac{\partial Q}{\partial A}}{w_{A}} > \frac{\bar{p} . \frac{\partial Q}{\partial L}}{w_{L}}

(1)

where $\bar{p}$ is the price of the produced commodity Q.

In this scenario, firms will hire human workers instead of androids only if the wage rate $w_{L}$ decreases enough, or human productivity increases enough, or androids’ productivity decreases enough, or androids’ remunerationincreases enough. However, if androids are humanoid machines capable of performing the same tasks as human workers, and the stock of other productive factors (e.g., the size of the factory) remains unchanged, their productivity is likely to move together. This means the productivity gap remains the same even if the use of androids increases and that of human workers decreases. As a result, firms may be willing to pay human workers a lower wage than they pay androids to compensate for this productivity difference (see the appendix for a more detailed analysis). However, in the AI era, the productivity gap between androids and human workers could become significant, leading to the possibility that the wage rate human workers need to accept to be hired could easily fall below subsistence levels—and consequently below all human workers’ reservation wages. The fact that human workers’ reservation wages cannot fall below subsistence can easily lead to full unemployment, since no worker would accept a wage below subsistence.

Clearly, for this situation to be effectively realized in practice, not simply as a long-term tendency, a series of conditions must occur, including effective technical substitutability, social and political feasibility, and the absence of frictions and exogenous shocks of various kinds, including public policies. We can therefore consider full unemployment as a situation toward which the system would asymptotically tend only in the long run.

3. Economists, AI, and Technological Unemployment

Curiously enough, the possibility that mass technological unemployment might occur has been at the very center of economists’ interest in the (remote and less remote) past, when the impact of technological progress on long-term unemployment ultimately proved to be mild empirically and almost non-existent theoretically. And it seems not to be at the very center of their interest today, when at least theoretical (even if not yet empirical) reasons appear much more robust in hypothesizing a massive negative impact of technological progress on employment. Only partially different has been the fate of the notion of “full unemployment,” which was neither discussed nor mentioned by economists in the past, when the state of technology (at that time and foreseeable) made its realization not only impossible but also impossible to forecast; and that even today remains a concept that is (almost) utterly absent in economic journals, not even mentioned as a situation that cannot occur, even though the interests of many social scientists other than economists, together with theoretical considerations, might suggest putting the topic center stage.² A further confirmation of economists’ attitude can be found in the circumstance that those (few) theoretical contributions describing scenarios similar to full unemployment, such as Korinek and Juelfs (2022) and D’Orlando (2024), can be found in personal repositories, working papers, or books—places where reviewers’ controls over publication are less strict.

Three appear to be the most likely explanations for this economists’ attitude: (i) the overvaluation of the importance of empirical studies, connected to the “dictatorship” of observable data, according to which a phenomenon without empirical confirmation does not deserve theoretical interest; and (ii) the aprioristic assumption that the substitution rate among human workers and machines will never be equal to one or close to one (or at least that human workers’ wage-to-productivity ratio will not be higher than machines’ remuneration-to-productivity ratio) in a relevant number of industries, so that an old theory, which Marx called “compensation theory,” which proved to be valid for the last three industrial revolutions, will also be valid for the Fourth; (iii) finally, the circumstance that mainstream economic approach faces significant difficulties in dealing with scenarios in which income distribution is driven by class struggle rather than factor endowment and marginal productivity, the market equilibrium wage drops below subsistence, and both full employment and free market dynamics lose relevance.

Let us begin by discussing the role of empirical studies.

3.1. The overstated relevance of empirical evidence

It is a fact that today’s empirical evidence does not support the claim that AI diffusion leads to mass unemployment. Several empirical studies discussing, in general, the impact of new technologies on employment exist, and most of them conclude that we can have waves of sectoral unemployment and wage reduction for some categories of workers, with the rise of income inequality (see, e.g., Dustmann et al. 2009; Graetz and Michaels 2018; Acemoglu and Restrepo 2020), but mass technological unemployment seems to be out of the question³ and “the fear of a jobless future. . . may be exaggerated and lacks an empirical base” (Hötte et al. 2023: 16), “or at the very least lacks empirical support so far” (Guarascio et al., 2024: 4). When AI is explicitly considered, these results appear empirically confirmed regarding inequality and unemployment, even if some authors are less optimistic about the impact on unemployment for certain groups of workers, tasks, or regions (see, e.g., Bordot 2022; Acemoglu 2024; Demirci et al. 2023; Hui et al. 2023; You et al. 2024).

In the past, during the “golden age” of economic theory, the absence of empirical evidence for a phenomenon that might or might not occur in the future would not have prevented economists from developing theoretical models to study it. But something changed at the sunset of the last century.

In the 1970s, we saw the birth, and later the diffusion, of a radical version of the neoclassical approach, with the development of models strictly based on the maximizing behavior of fully rational and fully informed subjects (or of subjects possessing rational expectations), on the deductive method, on strong mathematical formalizations, and on “ad hoc” simplifying assumptions. As a consequence, economics gained analytical strength but lost realism, suffering from the accusation of being incapable of describing actual reality and actual subjects’ behaviors. To reconnect economic theory with economic reality, economists began to extensively use new tools, like econometrics and experimental economics, to test the empirical relevance of their theoretical models and conclusions. Therefore, economic theory, and in particular (but not exclusively) the neoclassical approach, evolved from a mainly theoretical to a largely empirical discipline, so that if empirical evidence is lacking, the phenomenon itself does not deserve theoretical consideration. This has been the fate of full unemployment (and mass unemployment): Since today’s empirical data do not show high levels of unemployment, full and mass unemployment do not merit theoretical interest.

3.2. The theory of compensation

As for the second possible motivation for economists’ attitude, that is, compensation theory, we are referring to a theory whose initial development traces back to the first industrial revolutions. The First and Second Industrial Revolutions (1760–1900 and 1900–1960), by raising labor productivity, reduced labor requirements for a given amount of production in some sectors (to begin with, agriculture; later, manufacturing).⁴ Industrial revolutions were therefore thought to carry the potential side effect of lowering labor demand for the social categories involved. In such contexts, it was inevitable that economists questioned whether technological progress could raise long-term aggregate unemployment. However, aside from some caution among classical economists, most scholars argued that technological unemployment was not a concern because of the operation of compensation theory.

The compensation theory is based on the implicit (and, at the time of its birth, highly realistic) assumption that complete substitution of machines for human workers is impossible, and that mass substitution is unlikely in many industries. On these bases, the theory describes several mechanisms capable of returning the system to full employment once technological progress has generated (sectoral) unemployment. These mechanisms can be either automatic or deliberate, with automatic mechanisms mainly rooted in the neoclassical tradition and deliberate mechanisms in the Keynesian tradition.

The most important among the automatic mechanisms described in economic literature are the following: ⁵

i. Wage downward flexibility: In times of low employment, workers may accept lower wages, and firms may hire them at these reduced wages.

ii. Increase in demand for machines: The increased demand for machines caused by mechanization raises employment in the sector that produces machines.

iii. Increase in demand for goods due to price reduction: The mechanization of production reduces production costs, leading to lower prices for goods. This increases demand, boosts production, and creates more jobs, offsetting job losses caused by mechanization.

iv. Increase in labor and decrease in machines’ productivity: According to the mainstream approach, an increase in the use of production factors other than labor, such as capital, reduces these factors’ marginal productivity and raises labor’s marginal productivity, favoring techniques that require higher human employment.

v. New job opportunities: New technologies can create new job opportunities, offsetting job displacement caused by technological progress. In some recent contributions (see, e.g., Gregory et al. 2019; Dosi et al. 2022), this argument has been translated into the prevalence of the reinstatement effect, that is, the creation of new tasks for humans driven by technological progress, over the displacement effect, that is, a reduction in employment for tasks in which machines can substitute for humans.

The principal deliberate intervention mechanisms are the following:

vi. Increase in public expenditure. By definition, $P r o d u c t i v i t y = \frac{P r o d u c t i o n}{E m p l o y m e n t}$ and therefore

$E m p l o y m e n t = \frac{P r o d u c t i o n}{P r o d u c t i v i t y}$ , so that increased public expenditure, by generating increased aggregate demand and production, can offset reduced employment caused by increased productivity from mechanization.

vii. Luddite policies: Public policies may impose minimum labor hiring, limits on machine use, machine usage taxes, or worker hiring subsidies.

viii. Reduction in per-capita working hours: Reducing per-capita working hours can offset the need for fewer human workers for a given production.

At the beginning of the twenty-first century, in the years of the Third Industrial Revolution, with the diffusion of ICT technologies and robotization, compensation theory has been integrated by several contributions (often built within a neoclassical framework, with market clearing and overlapping generations) that study from a logical or formalized theoretical viewpoint the impact of automation on growth, wages, inequality (see, e.g., Violante 2008; Acemoglu and Autor 2011; Sachs and Kotlikoff 2012; Benzell et al. 2015; Sachs et al. 2015; Acemoglu and Restrepo 2018; Gasteiger and Prettner 2020), and less often (because of the market-clearing assumption present in most of the models) unemployment (see, e.g., Acemoglu and Restrepo 2019, 2020; Dosi et al. 2022). The most recent contributions explicitly consider AI (see, e.g., Berg et al. 2018; Korinek 2019; Korinek and Stiglitz 2019; Korinek and Stiglitz 2021; Korinek and Suh 2024; Acemoglu 2024; Harit 2024; Korinek and Trammell 2025).

The main conclusions of these new approaches align with those of the “old” compensation theory and “new” empirical studies: Technological progress can seriously risk increasing wage polarization and income inequality, but it does not substantially impact unemployment. Furthermore, with reference to AI diffusion, among the most relevant contributions only Korinek and Juelfs (2022) consider a full unemployment scenario in their third concern, that is, strong economic redundancy of labor. Berg et al. (2018) and Korinek’s (2019) present merely benchmark scenarios that they consider implausible.

We can conclude that numerous empirical and theoretical contributions support the “consensus view” that technological progress is unlikely to cause widespread long-term unemployment or, even more so, full unemployment, even after AI becomes widespread. Consequently, according to the consensus view, these ideas do not warrant further specific theoretical investigation.

3.3. The weaknesses of the mainstream approach

A third possible explanation for economists’ attitude is more specific and concerns how the mainstream (i.e., neoclassical) approach is structured and the economic aspects of reality it can address. In fact, the theory encounters significant problems when analyzing topics such as unemployment in general and, more specifically, mass unemployment, not to mention full unemployment.

The first issue arises from the same definition of equilibrium in the labor market, both for unemployment in general and full unemployment in particular. Regarding unemployment in general, the assumption of market-clearing dynamics is particularly problematic because it prevents the theory from fully capturing the many aspects of unemployment. Indeed, assuming market-clearing dynamics places all the theoretical emphasis on full-employment equilibria, viewed as stable positions toward which the economic system tends to move. In these equilibria, all individuals willing to work at the equilibrium wage are hired, while others remain voluntarily unemployed, thereby downplaying involuntary unemployment. Regarding full unemployment, defined as a situation where no workers accept employment because the equilibrium wage is below the subsistence level, the theory presents a rather paradoxical scenario: The economy would be in full unemployment (since no one would accept a job if the wage is below their survival needs) and in full employment (because all unemployed workers are voluntarily choosing not to work at the equilibrium wage rate). A conclusion that is difficult to accept.

Another shortcoming concerns income distribution. According to neoclassical economics, income distribution should be based on marginal productivity, which is determined mainly by factor endowments. On the contrary, in a full unemployment scenario, income distribution will no longer depend on factor endowments but on the bargaining power of social classes, since society will inevitably be divided into two conflicting classes: entrepreneurs (and/or android owners) and unemployed workers. Each social class will have different, and in most cases incompatible, goals. Class struggle will, therefore, be unavoidable and should be placed at the core of the theory in place of marginal productivity. Strictly linked to these considerations is the determination of wages. In a situation of full technological unemployment, equilibrium wages are by definition below subsistence levels, necessitating public policies that subsidize unemployed human workers, and/or subsidize firms for hiring human workers by paying them a wage above subsistence and above their marginal productivity, so that the traditional relationship between labor’s marginal productivity and real wages is disrupted. The maximizing procedure may still be valid (firms will, in any case, try to maximize profits under the new interventionist policy measures), even with a reduced scope (wages, labor employment, or android employment would be decided by policy, not by maximizing procedures), but the optimality of free market dynamics, the most important implication of the neoclassical approach, would be irremediably compromised.

Ultimately, it is the same General Equilibrium Approach on which conventional neoclassical models are founded that faces difficulties in dealing with full unemployment: The Walras-Arrow-Debreu-McKenzie general equilibrium framework can be adapted to discuss inequality, but it is unsuitable for discussing mass and, a fortiori, full unemployment as equilibrium positions. So, it is not coincidental that most contemporary versions of compensation theory focus on inequality rather than unemployment (see, e.g., Berg et al. 2018; Sachs and Kotlikoff 2012; Sachs et al. 2015).

Therefore, it is completely understandable, even if not justified, that neoclassical (mainstream) economists, lacking solid interpretive tools for addressing full unemployment, tend to omit this topic from what they consider deserving of theoretical interest.

More generally, all the motivations discussed in the three subsections above can help explain why economists (and journal reviewers) were skeptical that the diffusion of AI would have a significant impact on employment and, consequently, devoted little theoretical attention to (mass and) full unemployment.

4. Why the Economists’ Attitude Rests on Weak Foundations

We have observed that economists’ skepticism about the likelihood that AI will significantly impact employment, and their subsequent disinterest in proposing comprehensive theoretical studies on (mass and, a fortiori) full technological unemployment, are based on three points: (i) an overemphasis on the importance of empirical evidence; (ii) the a priori assumption that compensation theory will remain valid; and (iii) their dependence on an approach, namely the neoclassical economic model, that is poorly suited to address this issue.

We argue that both points (i) and (iii) cannot be valid justifications for economists’ attitude, and that point (ii) should be proven, but that proving it is impossible (and that, in attempting to do so, full unemployment equilibria need anyway the theoretical attention they deserve).

As usual, let us first discuss the role of empirical evidence.

4.1. The overstated relevance of empirical evidence

AI is just beginning to enter productive processes, and we should recognize that this entry is progressing differently across sectors and countries: The full impact of AI on employment will therefore only occur in the future, after it has fully spread. This suggests that current empirical data are useless for predicting the future impact of these new technologies on employment and, consequently, the chances of full or mass unemployment. Empirical evidence cannot, therefore, prove or disprove anything.

Even Hötte et al. (2023), who find ample support in the literature for the idea that “technological progress has not resulted in a negative net employment effect in the past decades” (Hötte et al. 2023: 16), maintain that:

[our] systematic review is subject to a number of limitations. First, empirical studies can only cover the impact of technologies already available today, but the scope of tasks that may be automated in the near future continuously expands.. . . Hence, empirical evidence on the impact of artificial intelligence, quantum computing, virtual reality, biotechnology, nanotechnology, renewable energy, and other emerging technologies that will soon impact our economy remains limited. None of the studies in our review assessed the impact of this new wave of technological innovation. To that end, it is unclear to which extent our findings can be extrapolated into the future. (Hötte et al. 2023: 19)

Similar considerations have been proposed by Thierer:

AI will cause job dislocations, of course, but no one can accurately predict which or how many jobs will be affected. Forecasting the future workforce is haunted by the same problem experts have always faced: We do not even possess a vocabulary to describe the jobs or skills of the future. (Thierer 2024: 4)

On the same line of thought is Saam:

The public debate often focuses on the job losses resulting from automation. Empirical evidence to date, however, does not point to any aggregate job losses.. . . Whether the current wave of technological progress based on machine learning technologies has different effects is yet to be seen. (Saam 2024: 23–24)

Overvaluing the relevance of today’s empirical data for discussing future phenomena probably played a crucial role in preventing economists from devoting the attention of study it deserves to the impact of AI on unemployment, but this attitude is fully unjustified.

Empirical confirmation of economic theory is undoubtedly a good thing. But it becomes a bad thing if such an attitude prevents economists from studying relevant phenomena for which empirical data are negligible or non-existent, as it inevitably is for phenomena that might (or might not) occur only in the future.

4.2. The theory of compensation

As noted above, the second possible explanation for economists’ attitude toward full unemployment relies on the continued validity of compensation theory. The validity of this theory rests on the a priori assumption that the substitution rate between human workers and machines will never be exactly one, or at least that human workers’ wage-to-productivity ratio will not exceed machines’ remuneration-to-productivity ratio, as empirical data seem to confirm. If this holds true, compensation theory—proven empirically valid through the last three industrial revolutions—will also apply to the Fourth without requiring a detailed, point-by-point theoretical analysis of the effectiveness of compensation forces in the new context.

But is the above assumption about the substitution rate still valid after introducing AI and machines endowed with AI into productive processes? Furthermore, can we assume its validity for future years as well?

We have already extensively discussed the main novelty of the Fourth Industrial Revolution: Nowadays, it is necessary to consider the possibility that machines endowed with AI can substitute for human workers in all jobs and tasks, not just increase their productivity. Therefore, we are in a completely new scenario: What was once inconceivable is now conceivable, at least as a long-term outcome, and in this new scenario, nothing prevents the substitution rate from being equal to one or close to one in most (or even all) industries. While the future remains uncertain, in section 2, we saw that recent studies on job exposure to AI indicate that approximately half of the workforce is already facing competition from AI, particularly in cognitive-task-oriented jobs. And this is the exposure of today’s jobs to today’s AI, with AI improving rapidly and expanding even faster into other sectors and jobs. It also seems difficult to imagine that the machines’ remuneration-to-productivity ratio will exceed that of human workers. But the theory of compensation, or at least most of its rebalancing mechanisms, is implicitly based on the assumption that the substitution rate cannot be equal to or close to one. So, either economists furnish theoretical proof that the substitution rate cannot be equal to one or close to one in the future (and/or that workers’ wage-to-productivity ratio cannot be higher than machines’ remuneration-to-productivity ratio), or they propose a point-by-point re-evaluation of the still capability of the different (automatic and/or deliberate) mechanisms hypothesized by compensation theory to play their rebalancing role in a world in which the substitution rate is equal to one or close to one. This is independent of empirical proof, since we have already seen that empirical evidence for phenomena that have not yet occurred is impossible to obtain.

Until today, no theoretical proof has been furnished that the substitution rate cannot be equal to or close to one (or that workers’ wage-to-productivity ratio cannot be higher than machines’ remuneration-to-productivity ratio), and such proof is probably impossible to be furnished. So, what remains is the evaluation of the capability of compensation forces to play their rebalancing role in the new reality. The problem is that at first glance, the theory of compensation does not appear to be well equipped for dealing with the new reality emerging from the Fourth Industrial Revolution, since most of the mechanisms that it describes seem inadequate for bringing the system back to full employment after an AI technological unemployment shock.

In particular, if in line with our thought experiment we admit that any increase in demand for goods (or services) can be met by increasing production by only using androids, without human contribution, then among the list of compensatory mechanisms presented in section 3.2, only wage downward flexibility, increased labor and decreased machines’ productivity, and Luddite policies seem to remain effective. And indeed, these mechanisms also face significant issues: For wage downward flexibility, nothing prevents wages from falling below the minimum subsistence level before human workers’ productivity-to-wage ratio exceeds that of androids, making downward wage flexibility ineffective for increasing employment; for workers’ and machines’ productivity, in a scenario where androids are humanoid machines capable of performing the same tasks as human workers, an increase in android use will decrease both human and android productivity at the same time. Only Luddite policies seem ultimately capable of surviving. But Luddite policies can prevent the system from reaching full unemployment; they do not suppress the spontaneous tendency of actual magnitudes to converge toward it.

The “old” theory of compensation appears, therefore, rather incapable of justifying the denial of the possible occurrence of full unemployment equilibria in the absence of appropriate intervention policies. One might wonder whether its “amended” contemporary version is better suited for this purpose.

The main problem with most of these recent contributions, which model the impact of technological progress on growth, wages, and inequality, is that, as we have already pointed out, they are rooted in the neoclassical framework with market-clearing equilibria. This framework, in general, faces relevant difficulties when called upon to systematize the AI scenario, as discussed in sub-section 4.3 below. However, in particular, it suffers from its market-clearing assumption, which prevents it from consistently studying involuntary unemployment equilibria.

We can easily conclude that if the substitution rate is equal to (or close to) one, the great majority of compensation forces presented above, particularly those based on automatic mechanisms, are relatively ineffective in counteracting the rise in unemployment caused by the diffusion of AI. Therefore, economists’ attitude toward studying mass and full unemployment cannot be justified based on the theory of compensation, old and new, unless a proof of impossibility for the substitution rate equal to one (or of the unlikelihood of a greater wage/productivity ratio for workers compared to machines) is furnished.

Again, nobody supports the idea that fully replacing human workers with AI-powered machines across most or all sectors of the economy is certain to happen; it might or might not occur; nobody knows for sure today. We only argue that, after AI diffusion, assuming the system might asymptotically converge (in the long run) to full unemployment has at least the same theoretical dignity as assuming it might converge to full employment. Economists, therefore, have the moral obligation to use the tools of economic theory to study “what if” the substitution rate between human workers and machines is equal to one or close to one in most industries, and full or at least mass technological unemployment tends to occur, since today such an outcome can no longer be ruled out. Especially not a priori.

4.3. The weaknesses of the mainstream approach

The last point discussed in section 3.3 above, that traditional theory is ill-suited for studying unemployment in general and full unemployment in particular, can certainly explain, but on the other hand cannot justify, economists’ attitude of not studying mass and full unemployment equilibria in depth, not even as a long-run outcome. If the theory is inadequate, it is better to dismiss or modify it rather than insist on ignoring changes in the technological and economic landscape simply because the dominant theory cannot explain them.

To modify the theory, valuable insights can come from the past. In particular, classical economists and classical-type theories seem well equipped to provide the theoretical tools needed to analyze full unemployment. First, classical and classical-type approaches are well suited to studying situations in which class struggle, rather than marginal productivity and factor endowments, determines income distribution. Second, the classical notion of subsistence wages closely resembles circumstances in which equilibrium wages decline below subsistence levels, necessitating public policies to subsidize workers so they can attain subsistence and/or to subsidize firms to hire workers and pay them a subsistence wage. Finally, the study of long-period positions and the way actual reality converges toward and gravitates around them has a long-standing tradition in both classical and classical-type economic theory (see, e.g., Garegnani 1976, 1988).

Somewhat paradoxically, AI diffusion may bring economic theorization back to the era of classical Political Economy, emphasizing the roles of social classes, class struggle, long-period positions, and redistribution of surplus. These topics will become increasingly relevant in the near future, particularly as we face the challenge of extracting surplus from capitalist (and/or android owners) and redistributing it to the unemployed. This redistribution is vital for human survival and maintaining demand for goods, but it is not easy to achieve politically, since those who need it, namely the unemployed, will have much less bargaining power than before, because they are no longer essential for production.

5. Conclusions: Is Full Unemployment Worth Theoretical Interest?

The robustness of economists’ (and journal reviewers’) attitude toward the possible occurrence of (mass or) full unemployment crucially rests upon the likelihood of the circumstance that AI (and/or machines endowed with AI) will be able to substitute for human workers in all (or at least in most) jobs and tasks at a lower remuneration-to-productivity ratio: If such a substitution has a positive probability of coming true in the foreseeable future, the two related concepts of mass and full unemployment undoubtedly deserve theoretical interest and today’s economists’ attitude is unjustified; on the contrary, if there is no possibility at all that this substitution will occur, the two concepts are deprived of theoretical interest and economists’ attitude is fully justified. So, we should answer a simple question: How realistic is a future substitution rate between human workers and humanoid machines equal to, or close to, one?

The problem is that this question is unanswerable. As of today, AI and machines endowed with AI seem capable of substituting for human workers in many productive processes, regardless of workers’ skill endowment and whether they perform routine or non-routine tasks, a scenario that was previously unrealistic (and unpredictable). We cannot know whether this process of substitution will go so far as to create (mass or even) full technological unemployment; what we know is that economic theory developed to study the past industrial revolutions, as well as today’s empirical evidence, is useless for forecasting future economic outcomes caused by AI future development. So, the question is unanswerable also for economists. What economists can do, however, is consider the possibility that, in the absence of obstacles or counteracting policies, and if AI continues to spread and improve without slowing, the economic system may experience mass unemployment in the short term and asymptotically converge toward a full unemployment equilibrium in the long term. The suggestion is therefore to focus on the worst-case scenario and, in view of the diminishing role of human workers in production, treat full unemployment as an attractor toward which actual reality converges, essentially as classical and classical-type approaches treat long-period positions. Reintroducing a classical perspective would also be beneficial because this approach is better equipped than traditional neoclassical models to address phenomena such as class struggle and subsistence wages, which might be plausible outcomes of the spread of AI. For study purposes, we should therefore consider replacing the long-term trend toward full employment with a trend toward full unemployment, or at least treat both scenarios as valid subjects for theoretical exploration. This would also facilitate the study of the system’s short-term dynamics, which are influenced not only by the gradual integration of AI into productive processes but also by the fact that this integration does not occur simultaneously across all sectors. Additionally, we should also consider the direct and indirect effects of changes in demand patterns. For this purpose, the literature on structural change initiated by Pasinetti (1981) could serve as a valuable starting point for opening a significant new line of research.

While it is impossible to determine today whether a full unemployment scenario will occur in the foreseeable future, it is undeniably true that full (and mass) technological unemployment caused by the spread of AI warrants theoretical interest. It is also better to analyze the worst-case scenario of full unemployment in advance, before it occurs, rather than studying it afterward. Conducting such an analysis in advance could help us identify the most effective policy tools to prevent or minimize economic harm, rather than reacting after the damage has been done.

Footnotes

Appendix

In section 2, we have seen that, given the produced quantities, firms will hire androids (A) instead of human workers (L) if:

(1)

\frac{\bar{p .} \frac{\partial Q}{\partial A}}{w_{A}} > \frac{\bar{p} . \frac{\partial Q}{\partial L}}{w_{L}}

and the fact that androids are humanoid machines that can substitute for human workers in all jobs and tasks also means that when androids’ marginal product varies (because of an increase or decrease in the number of androids entering the productive process), the same happens to human workers’ marginal product. The co-movement of human workers’ and androids’ productivity implies that if humans are less productive than androids, they will always remain so, regardless of how many androids and/or human workers are hired by firms. In these circumstances, firms will only hire human workers if $\frac{w_{L}}{w_{A}}$ lowers enough.

Under our assumption that androids have higher productivity than human workers, we can also write:

(2)

\bar{p} . \frac{\partial Q}{\partial A} > \bar{p} . \frac{\partial Q}{\partial L}

Or:

(3)

\bar{p} . \frac{\partial Q}{\partial A} = α . \bar{p} . \frac{\partial Q}{\partial L} (with α > 1) .

Therefore, the following condition must be met to ensure human employment:

(4)

w_{L} = \frac{w_{A}}{α}

so that:

(5)

\frac{\bar{p .} \frac{\partial Q}{\partial A}}{w_{A}} = \frac{\bar{p} . \frac{\partial Q}{\partial L}}{\frac{w_{A}}{α}}

And:

(6)

\frac{\bar{p .} \frac{\partial Q}{\partial A}}{w_{A}} = \frac{\bar{p} . \frac{\partial Q}{\partial L}}{w_{L}}

So, firms will hire human workers only if they accept a wage rate $w_{L} = \frac{w_{A}}{α}$ . But in the AI era, the value of $α$ might be huge, so that $\frac{w_{A}}{α}$ might result below subsistence and no worker will accept to be hired for a wage below subsistence.

Moving from the firm level to the labor market level, the easiest (and most conventional) way to discuss both the determination of the equilibrium remuneration of human workers and androids, and the amount of employment, is to assume that the remuneration of both human workers and androids is determined by a competitive market. In such a scenario, human workers’ labor supply LS can be considered as a function of $w_{L} = \frac{w}{α}$ (with w being the net amount of money firms pay to those who perform a specific task in a given period of time):

(7)

L S = f (\frac{w}{α})

(8)

A S = f (w)

with AS′ > 0.

By equating LS + AS to the total “labor” demand by firms, we can determine w*, w* and $w_{L} = \frac{w^{*}}{α}$ together with both human workers’ and androids’ employment level. The graphic representation of the resulting equilibrium is proposed in figure A1: LS curve represents labor supplied by workers, AS “labor” supplied by androids, LS + AS is the aggregate “labor” supply curve, LD+AD is aggregate “labor” demand (for firms, it is indifferent to hire workers or androids if the former accept a wage $w_{L} = \frac{w}{α}$ ), L* and A* are, respectively, labor and androids equilibrium levels of employment. If the equilibrium remuneration w* that equates “labor” demand LD+AD and “labor” supply LS+AS is such that human workers’ wage $w_{L}^{*} = \frac{w^{*}}{α}$ is below subsistence (i.e., $w^{*} < α . s u b s i s t e n c e$ ), competitive forces will push the system, or at least this sector of the economy, to a full-unemployment equilibrium E, in which no human worker and A* androids will be hired, since workers will not accept to be hired for a remuneration below subsistence.

Things change slightly in the case of mass (or close-to-full) unemployment equilibrium, represented in figure A2, a situation in which most sectors of the economy are in full unemployment equilibria, so that the graph of figure A1 can represent them, and other sectors are in mass unemployment equilibria, with an equilibrium wage w* higher than $α . s u b s i s t e n c e$ , (i.e., $\frac{w^{*}}{α} > s u b s i s t e n c e$ ) so that a certain number of workers accepts being hired. In figure A2, L* shows the number of human workers hired in equilibrium, and A* is the number of androids hired.

Acknowledgements

I wish to thank the reviewers and editors for their insightful comments. In particular, I am very thankful to Davide Gualerzi, Baris Güven, and Gary Mongiovi for their valuable suggestions and critiques of earlier versions of this work. The usual disclaimer applies.

ORCID iD

Fabio D’Orlando

Ethical Considerations

This article does not contain any studies with human or animal participants.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

1

Brynjolfsson and McAfee (2011, 2014), Ford (2015), and made seminal contributions to the novelty represented by the above-described substitution process. However, even these scholars did not fully understand AI’s true potential in those days: It was still too early.

2

In Scopus, we find no books or articles published in economics journals over the past twenty-five years (2001–2025) containing the term “full unemployment” in the title, and only one article and one book written by economists include the exact phrase “full unemployment” in their abstracts. Additionally, in Scopus, only eighteen articles published in economics journals in the last twenty-five years feature the exact term “technological unemployment” in their titles, and just fifty articles in economics journals have the exact term “technological unemployment” in their abstracts.

3

Hötte et al. (2023) reviewed 127 studies published between 1988 and 2021 finding “larger support for the labor-creating effect of technological change” (: 2) and concluding that “overall, our findings strongly suggest that technological progress has not resulted in a negative net employment effect in the past decades” (Hötte et al. 2023: 16).

4

The impact of the first two Industrial Revolutions on employment in the primary and secondary sectors is effectively described by Campa (2018: 59): “The transition from traditional agriculture to intensive agriculture, through the use of agricultural machinery, herbicides, fertilizers, fungicides, etc., has led to demographic emptying of the countryside. The evaporation of jobs in the primary sector of the United States of America offers impressive numbers: in 1900 41 percent of the population was employed in agriculture. A century later, in 2000, only 2 percent of Americans still worked in same sector.. . . A similar phenomenon was observed in the secondary sector, or manufacturing, at the turn of the twentieth and twenty-first century. In the United States, the ratio between employment in the factories decreased from 22.5 percent in 1980 to 10 percent today and is expected further decline to below 3 percent by 2030.” On the impact of the first industrial revolutions on sectoral employment, see also Schettkat and Yocarini (2003) and Campa (2014, ).

5

For a list and discussion of compensation mechanisms, see Vivarelli (2007), Blien and Ludewig (2017), Campa (2018), Peters (2017), D’Orlando (2020, ).

Author Biography

Fabio D’Orlando is Associate Professor of Economics at the Department of Economics and Law, University of Cassino and Southern Lazio, Italy. He studied at Sapienza University of Rome, Italy (PhD), and later taught there and at the University of Campobasso, Italy. He currently teaches Economics and Monetary Economics at the University of Cassino. Additionally, he teaches economic policy, commercial, financial, and multilateral cooperation for the master’s program in Diplomatic Studies at the Italian Society for International Organization (SIOI) in Rome. He leads the Economic Sciences Section, serves on the teaching board for the PhD program in Economics and Management for Innovation and Sustainability, and acts as the scientific coordinator for the University of Cassino on the European TNE-IMPACT PNRR research project. His research interests are technological unemployment, behavioral economics, history of economic ideas, and classical-type theory. For Routledge, he is the author of Technological Unemployment, Basic Income, and Wellbeing (2024) and (co-)author of Economic Change and Wellbeing. The True Cost of Creative Destruction and Globalization (2022), written with Francesco Ferrante and Albertina Oliverio.

References

Acemoglu

Daron

. 2024. Automation, AI, and wages. In Navigating the Future of Work. Perspectives on Automation, AI, and Economic Prosperity, eds. Erik

Brynjolfsson

Thierer

Adam

Acemoglu

Daron

. Washington, DC: American Enterprise Institute.

Acemoglu

Daron

Autor

David

. 2011. Skills, tasks, and technologies: Implications for employment and earnings. In Handbook of Labor Economics, eds. Orley

Ashenfelter

Card

David

, volume 4, part B. San Diego, CA: North Holland.

Acemoglu

Daron

Johnson

Simon

. 2023. Rebalancing AI. F&D Magazine, December: 26–29.

Acemoglu

Daron

Restrepo

Pascual

. 2018. The race between machines and humans: Implications for growth, factors shares and jobs. American Economic Review 108 (6): 1488–542.

Acemoglu

Daron

Restrepo

Pascual

. 2019. Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives 33 (2): 3–30.

Acemoglu

Daron

Restrepo

Pascual

. 2020. Robots and jobs: Evidence from US labor markets. Journal of Political Economy 128 (6): 2188–244.

Ali

Sabah Hameed

Ayad

Hayder

Al Rubaie

Mithaq Taher

. 2022. Fifth industrial revolution: (new perspectives). International Journal of Business, Management and Economics 3 (3): 196–212.

Benzell

Seth

Kotlikoff

Laurence J.

LaGarda

Guillermo

Sachs

Jeffrey D.

2015. Robots Are Us: Some Economics of Human Replacement. NBER Working Paper no. 20941. Cambridge, MA: National Bureau of Economic Research. Accessed at: http://www.nber.org/papers/w20941.

Berg

Andrew

Buffie

Edward

Zanna, L

Luis-Felipe

. 2018. Should we fear the robot revolution? (The correct answer is yes). Journal of Monetary Economics 97 (C): 117–48.

10.

Blien

Uwe

Ludewig

Oliver

. 2017. Technological Progress and (Un)employment Development. IZA Discussion Paper No. 10472. Bonn, Germany: IZA Institute. Accessed at: http://ftp.iza.org/dp10472.pdf.

11.

Bordot

Florent

. 2022. Artificial Intelligence, robots, and unemployment: Evidence from OECD countries. Journal of Innovation Economics & Management 37: 117–38.

12.

Brynjolfsson

Erik

McAfee

Andrew

. 2011. Race Against the Machine. Lexington: Digital Frontier Press.

13.

Brynjolfsson

Erik

McAfee

Andrew

. 2014. The Second Machine Age. New York: W. W. Norton.

14.

Campa

Riccardo

. 2014. Workers and automata: A sociological analysis of the Italian case. Journal of Evolution and Technology 24 (1): 70–85.

15.

Campa

Riccardo

. 2018. Technological unemployment: A brief history of an idea. Orbis Idearum 6 (2): 57–79.

16.

Cazzaniga

Mauro

Jaumotte

Florence

Longji

Melina

Giovanni

Panton

Augustus

Pizzinelli

Carlo

Rockall

Emma

Tavares

Marina Mendes

. 2024. Gen-AI: Artificial Intelligence and the Future of Work. IMF Staff Discussion Note SDN2024/001. Washington, DC: International Monetary Fund. Accessed at: https://www.imf.org/-/media/files/publications/sdn/2024/english/sdnea2024001.pdf.

17.

Dator

Jim

. 2016. Grands Enjeux du Monde de Demain: Robotics and Artificial Intelligence. Presentation at Institut Royal des Etudes Stratégiques, Bibliotheque nationale du Royaume du Maroc, July 13. Accessed at: https://www.ires.ma/images/pdfs/Forums/Conf%C3%A9rences/DatorIRES-July-13.pdf.

18.

Dator

Jim

. 2020. Editorial: Thoughts on Big Data, AI, and futures studies. World Futures Review 12 (2): 151–55.

19.

Demirci

Ozge

Hannane

Jonas

Zhu

Xinrong

. 2023. Who Is AI Replacing? The Impact of Generative AI on Online Freelancing Platforms. Accessed at: https://ssrn.com/abstract=4602944 or http://dx.doi.org/10.2139/ssrn.4602944.

20.

D’Orlando

Fabio

. 2020. Technological unemployment and the resurgence of political economy. American Review of Political Economy 15 (1): 1–25.

21.

D’Orlando

Fabio

. 2024. Technological Unemployment, Basic Income, and Well-being. London: Routledge.

22.

D’Orlando

Fabio

Ferrante

Francesco

. 2022. Economic Change and Well-being. The True Cost of Creative Destruction and Globalization. London: Routledge.

23.

Dosi

Giovanni

Pereira

Marcelo

Roventini

Andrea

Virgillito

Maria Enrica

. 2022. Technological paradigms, labour creation, and destruction in a multi-sector agent-based model. Research Policy 51 (10): 104565.

24.

Dustmann

Christian

Ludsteck

Johannes

Schönberg

Uta

. 2009. Revisiting the German wage structure. The Quarterly Journal of Economics 124 (2): 843–81.

25.

Eloundou

Tyna

Manning

Sam

Mishkin

Pamela

Rock

Daniel

. 2024. GPTs are GPTs: labor market impact potential of LLMs. Science 384 (6702): 1306–1308.

26.

Ford

Martin

. 2015. Rise of the Robots. Technology and the Threat of a Jobless Future. New York: Basic Books.

27.

Garegnani

Pierangelo

. 1976. On a change in the notion of equilibrium in recent work on value and distribution. In Essays in Modern Capital Theory, eds. Murray

Brown

Sato

Kazuo

Zaremka

Paul

. Amsterdam: North-Holland.

28.

Garegnani

Pierangelo

. 1988. Actual and normal magnitudes: A comment on Asimakopulos. Political Economy, Studies in the Surplus Approach 4 (2): 251–58.

29.

Gasteiger

Emanuel

Prettner

Klaus

. 2020. Automation, Stagnation, and the Implications of a Robot Tax. ECON WPS Working Papers in Economic Theory and Policy no. 02/2020. Wien: Institute of Statistics and Mathematical Methods in Economics, Economics Research Unit. Accessed at: https://EconPapers.repec.org/RePEc:zbw:tuweco:022020.

30.

Graetz

Georg

Michaels

Guy

. 2018. Robots at Work. IZA Discussion Paper no. 8938. Bonn, Germany: IZA Institute. Accessed at: https://docs.iza.org/dp8938.pdf.

31.

Gregory

Terry

Salomons

Anna

Zierahn

Ulrich

. 2019. Racing With or Against Machine? Evidence from Europe. IZA Discussion Paper no. 12063. Bonn, Germany: IZA Institute. Accessed at: https://docs.iza.org/dp12063.pdf.

32.

Guarascio

Dario

Piccirillo

Alessandro

Reljic

Jelena

. 2024. Will Robot Replace Workers? Assessing the Impact of Robots on Employment and Wages with Meta-analysis. LEM Working Paper Series no. 2024/03. Pisa, Italy: Institute of Economics, Scuola Superiore Sant’Anna. Accessed at: doi:10.13140/RG.2.2.31700.58246.

33.

Harit

Aditya

. 2024. The Economic Implications of AI-Driven Automation: A Dynamic General Equilibrium Analysis. MPRA Paper no. 122244. Munich: Munich Personal RePEc Archive. Accessed at: https://mpra.ub.uni-muenchen.de/122244/1/MPRA_paper_122244.pdf.

34.

Hötte

Kerstin

Somers

Melline

Theodorakopoulos

Angelos

. 2023. Technology and jobs: A systematic literature review. Technological Forecasting and Social Change 194 (122750): 1–23.

35.

Hui

Xiang

Reshef

Oren

Zhou

Luofeng

. 2023. The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market. Accessed at: https://ssrn.com/abstract=4527336 or https://doi.org/10.2139/ssrn.4527336.

36.

Korinek

Anton

. 2019. The Rise of Artificially Intelligent Agents. Accessed at: https://drive.google.com/file/d/16y5UmeTOv5YB9E5ms_ce7WiYNfMAn17J/view.

37.

Korinek

Anton

Juelfs

Megan

. 2022 Preparing for the (Non-Existent?) Future of Work. NBER Working Paper no. 30172. Cambridge, MA: National Bureau of Economic Research. Accessed at: http://www.nber.org/papers/w30172.

38.

Korinek

Anton

Stiglitz

Joseph

. 2019. AI, Worker-Replacing Progress, and the Coming Singularity. Accessed at: https://drive.google.com/file/d/1Xuq9hiZWCzmbxilosh7578FDhwO49AkJ/view.

39.

Korinek

Anton

Stiglitz

Joseph

. 2021. Steering Technological Progress. Accessed at: https://www.dropbox.com/scl/fi/asbmis148m9v6xy4i7rd5/SteeringProgressV6.pdf?rlkey=hdmzjfwth6w2s5y8b90g4stv5&e=1&dl=0.

40.

Korinek

Anton

Suh

Donghyun

. 2024. Scenarios for the Transition to AGI. Accessed at: https://www.dropbox.com/scl/fi/viob7f5yv13zy0ziezlcg/AGI_Scenarios.pdf?rlkey=8hxq9rm82kksocw1zjilcxf8v&e=1&dl=0.

41.

Korinek

Anton

Trammel

Philip

. 2025. Economic Growth under Transformative AI. NBER Working Paper no. 31815. Cambridge, MA: National Bureau of Economic Research. Accessed at: http://www.nber.org/papers/w31815.

42.

Luo

Jianxi

. 2023. Designing the future of the fourth industrial revolution. Journal of Engineering Design 34 (10): 779–85.

43.

Morrar

Rabeh

Husma

Arman

Mousa

Saeed

. 2017. The fourth industrial revolution (Industry 4.0): A social innovation perspective. Technology Innovation Management Review 7 (11): 12–20.

44.

Mourtzis

Dimitris

. 2021. Towards the 5th industrial revolution: A literature review and a framework for process optimization on big data analytics and semantic. Journal of Machine Engineering 21 (3): 5–39.

45.

Muir

George

. 2018. AI—the fifth industrial revolution. AI Business. Accessed at: https://aibusiness.com/verticals/ai-the-fifth-industrial-revolution.

46.

Pasinetti

Luigi Ludovico

. 1981. Structural Change and Economic Growth. Cambridge University Press.

47.

Peters

Michael Adrian

. 2017. Technological unemployment: Educating for the fourth industrial revolution. Educational Philosophy and Theory 49 (1): 1–6.

48.

Saam

Marianne

. 2024. The impact of artificial intelligence on productivity and employment—How can we assess it and what can we observe? Intereconomics 59 (1): 22–27.

49.

Sachs

Jeffrey

Kotlikoff

Laurence

. 2012. Smart Machines and Long-term Misery. NBER Working Paper no. 18629. Cambridge, MA: National Bureau of Economic Research. Accessed at: https://www.nber.org/papers/w18629.

50.

Sachs

Jeffrey

Benzell

Seth

LaGarda

Guillermo

. 2015. Robots: Curse or Blessing? A Basic Framework. NBER Working Paper no. 21091. Cambridge, MA: National Bureau of Economic Research. Accessed at: http://www.nber.org/papers/w21091.

51.

Schettkat

Ronald

Yocarini

Lara

. 2003. The Shift to Services. A Review of the Literature. IZA Discussion Paper no. 964. Bonn, Germany: IZA Institute. Accessed at: http://ftp.iza.org/dp964.pdf.

52.

Schwab

Klaus

. 2016. The Fourth Industrial Revolution. Geneve: World Economic Forum.

53.

Thierer

Adam

. 2024. We can’t predict the future of work. In Navigating the Future of Work. Perspectives on Automation, AI, and Economic Prosperity, eds. Erik

Brynjolfsson

Thierer

Adam

Acemoglu

Daron

. Washington, DC: American Enterprise Institute.

54.

Violante

Giovanni

. 2008. Skill-biased technical change. In The New Palgrave Dictionary of Economics, eds. Matias

Vernengo

Caldentey

Esteban Perez

Ghosh

Jayati

. London: Palgrave Macmillan.

55.

Vivarelli

Marco

. 2007. Innovation and Employment: A Survey. IZA Discussion Paper no. 2621. Bonn, Germany: IZA Institute. Accessed at: http://ftp.iza.org/dp2621.pdf.

56.

Webb

Michael

. 2020. The Impact of Artificial Intelligence on the Labor Market. SSRN Working Paper, January 11. Accessed at: https://ssrn.com/abstract=3482150 or https://doi.org/10.2139/ssrn.3482150.

57.

West

Darrell

. 2015. What Happens if Robots Take the Jobs? The Impact of Emerging Technologies on Employment and Public Policy. Center for Technology Innovation at Brookings, October 26. Washington, DC: The Brookings Institution. Accessed at: https://www.brookings.edu/wp-content/uploads/2016/06/robotwork.pdf.

58.

Min

David

Jeanne

Kim

Suk Hi

. 2018. The fourth industrial revolution: Opportunities and challenges. International Journal of Financial Research 9 (2): 90–95.

59.

You

Jing

Xiangyu

Liao

Deng

Lin

Chen

. 2024. International comparison of the impact of digital transformation on employment. Journal of Asian Economics 95: 101820.